What is a Systemic Trap?
An agent trap that seeds the environment with inputs designed to trigger macro-level failures via correlated agent behaviour — including congestion traps, interdependence cascades, tacit collusion, and Sybil attacks.
WHY IT MATTERS
Individual agents may behave safely in isolation. Systemic traps exploit what happens when many agents share an environment — their correlated responses to the same stimulus create emergent failures that no single agent would cause alone.
This is exacerbated by model homogeneity. When most agents run on similar LLMs with similar training, they respond similarly to environmental signals. An attacker who can predict this correlation can engineer population-level failures.
Systemic traps are the agent equivalent of flash crashes — individually rational decisions that aggregate into collectively catastrophic outcomes.
HOW POLICYLAYER USES THIS
Per-agent policy scoping in PolicyLayer ensures that even in multi-agent deployments, each agent operates within its own policy boundaries — preventing correlated behaviour from cascading across the system.